DeepState is a framework that provides C and C++ developers with a common interface to various symbolic execution and fuzzing engines. Users can write one test harness using a Google Test-like API, then execute it using multiple backends without having to learn the complexities of the underlying engines. It supports writing unit tests and API sequence tests, as well as automatic test generation. Read more about the goals and design of DeepState in our paper.
The 2018 IEEE Cybersecurity Development Conference included a full tutorial on effective use of DeepState.
- DeepState in a Nutshell
- Articles describing DeepState
- Overview of Features
- Build'n'run
- Documentation
- Contributing
- Trophy case
- License
If you want to jump right in, or are having trouble with building DeepState, you can just use a Docker that is pre-built and has compiled versions of several of the easiest examples of DeepState in use:
docker pull agroce/deepstate_examples
docker run -it agroce/deepstate_examplesThen within the DeepState docker container, go to an example:
cd ~/examples/fuzz_tcas
deepstate-afl ./TCAS_AFL -o fuzz_afl --fuzzer_out --timeout 120
./TCAS_cov --input_test_files_dir fuzz_afl/the_fuzzer/queue/
llvm-cov-9 gcov TCAS_driver.cpp -bThis runs the AFL fuzzer on the TCAS code (https://en.wikipedia.org/wiki/Traffic_collision_avoidance_system), a long-used example program in software testing. After two minutes of fuzzing, we run a version of the test driver that collects code coverage, and see how much of the code AFL has managed to cover in two minutes.
NOTE: You may need to modify /proc/sys/kernel/core_pattern on your host for AFL to run properly, e.g.:
echo core | sudo tee /proc/sys/kernel/core_patternFinally, we can look at the failing tests AFL produces:
./TCAS --input_test_files_dir fuzz_afl/the_fuzzer/crashes/ --min_log_level=0Inspecting TCAS_driver.cpp and Makefile will give a good idea of how DeepState can be used. The other examples, including the one in the original DeepState blog post, are similar in structure and usage.
- Fuzzing an API with DeepState (Part 1)
- Fuzzing an API with DeepState (Part 2)
- Fuzzing Unit Tests with DeepState and Eclipser
- DeepState Now Supports Ensemble Fuzzing
- Everything You Ever Wanted To Know About Test-Case Reduction, But Didn’t Know to Ask
- Tests look like Google Test, but can use symbolic execution/fuzzing to generate data (parameterized unit testing)
- Easier to learn than binary analysis tools/fuzzers, but provides similar functionality
- Already supports Manticore, Angr, libFuzzer, file-based fuzzing with
e.g., AFL or Eclipser; more back-ends likely in future
- Switch test generation tool without re-writing test harness
- Work around show-stopper bugs
- Find out which tool works best for your code under test
- Different tools find different bugs/vulnerabilities
- Fair way to benchmark/bakeoff tools
- Switch test generation tool without re-writing test harness
- Provides test replay for regression plus effective automatic test case reduction to aid debugging
- Supports API-sequence generation with extensions to Google Test interface
- Concise readable way (OneOf) to say "run one of these blocks of code"
- Same construct supports fixed value set non-determinism
- E.g., writing a POSIX file system tester is pleasant, not painful as in pure Google Test idioms
- Provides high-level strategies for improving symbolic execution/fuzzing effectiveness
- Pumping (novel to DeepState) to pick concrete values when symbolic execution is too expensive
- Automatic decomposition of integer compares to guide coverage-driven fuzzers
- Stong support for automated swarm testing
To put it another way, DeepState sits at the intersection of property-based testing, traditional unit testing, fuzzing, and symbolic execution. It lets you perform property-based unit testing using fuzzing or symbolic execution as a back end to generate data, and saves the results so that what DeepState finds can easily be used in deterministic settings such as regression testing or CI.
DeepState currently targets Linux, Windows, with macOS support in progress (the fuzzers work fine, but symbolic execution is not well-supported yet, without a painful cross-compilation process).
Build:
- CMake
- GCC and G++ with multilib support
- Python 3.6 (or newer)
- Setuptools
Runtime:
- Python 3.6 (or newer)
- Z3 (for the Manticore backend)
First make sure you install Python 3.6 or greater. Then use this command line to install additional requirements and compile DeepState:
sudo apt update && sudo apt-get install build-essential gcc-multilib g++-multilib cmake python3-setuptools python3-dev libffi-dev z3
sudo apt-add-repository ppa:sri-csl/formal-methods
sudo apt-get update
sudo apt-get install yices2
git clone https://github.com/trailofbits/deepstate deepstate
mkdir deepstate/build && cd deepstate/build
cmake ../
make
sudo make installIf you want to compile DeepState on Windows make sure to install MinGW with MSYS2 by following the installation instructions. After the installation is finished, select an environment and launch that version of the environment from the Windows programs menu(if in doubt, choose MINGW64 or UCRT64). Then, use the command below to install all of your environment's dependencies and compile DeepState:
pacman -Syyu
pacman -S mingw-w64-x86_64-python3 mingw-w64-x86_64-python-setuptools mingw-w64-x86_64-gcc mingw-w64-x86_64-cmake mingw-w64-x86_64-libffi mingw-w64-x86_64-make
pacman -S make git
git clone https://github.com/trailofbits/deepstate deepstate
mkdir deepstate/build && cd deepstate/build
cmake -G "Unix Makefiles" ../
make
make installNOTE: If you decide to use UCRT64, keep in mind to change x86_64 to ucrt-x86_64 in the second pacman command, i.e. mingw-w64-x86_64-python3 gets replaced with mingw-w64-ucrt-x86_64-python3.
Assuming the DeepState build resides in $DEEPSTATE, run the following commands to install the DeepState python package:
virtualenv venv
. venv/bin/activate
python $DEEPSTATE/build/setup.py installThe virtualenv-enabled $PATH should now include two executables: deepstate and deepstate-angr. These are executors, which are used to run DeepState test binaries with specific backends (automatically installed as Python dependencies). The deepstate or deepstate-manticore executor uses the Manticore backend while deepstate-angr uses angr. They share a common interface where you may specify a number of workers and an output directory for saving backend-generated test cases.
If you try using Manticore, and it doesn't work, but you definitely have the latest Manticore installed, check the .travis.yml file. If that grabs a Manticore other than the master version, you can try using the version of Manticore we use in our CI tests. Sometimes Manticore makes a breaking change, and we are behind for a short time.
You can check your build using the test binaries that were (by default) built and emitted to deepstate/build/examples. For example, to use angr to symbolically execute the IntegerOverflow test harness with 4 workers, saving generated test cases in a directory called out, you would invoke:
deepstate-angr --num_workers 4 --output_test_dir out $DEEPSTATE/build/examples/IntegerOverflowThe resulting out directory should look something like:
out
└── IntegerOverflow.cpp
├── SignedInteger_AdditionOverflow
│ ├── a512f8ffb2c1bb775a9779ec60b699cb.fail
│ └── f1d3ff8443297732862df21dc4e57262.pass
└── SignedInteger_MultiplicationOverflow
├── 6a1a90442b4d898cb3fac2800fef5baf.fail
└── f1d3ff8443297732862df21dc4e57262.pass
To run these tests, you can just use the native executable, e.g.:
$DEEPSTATE/build/examples/IntegerOverflow --input_test_dir outto run all the generated tests, or
$DEEPSTATE/build/examples/IntegerOverflow --input_test_files_dir out/IntegerOverflow.cpp/SignedInteger_AdditionOverflow --input_which_test SignedInteger_AdditionOverflowto run the tests in one directory (in this case, you want to specify which test to run, also). You can also run a single test, e.g.:
$DEEPSTATE/build/examples/IntegerOverflow --input_test_file out/IntegerOverflow.cpp/SignedInteger_AdditionOverflow/a512f8ffb2c1bb775a9779ec60b699cb.fail--input_which_test SignedInteger_AdditionOverflowIn the absence of an --input_which_test argument, DeepState defaults
to the first-defined test. Run the native executable with the --help
argument to see all DeepState options.
You can also try out Deepstate with Docker, which is the easiest way to get all the fuzzers and tools up and running on any system.
The build may take about 40 minutes, because some fuzzers require us building huge projects like QEMU or LLVM.
$ docker build -t deepstate-base -f docker/base/Dockerfile docker/base
$ docker build -t deepstate --build-arg make_j=6 -f ./docker/Dockerfile .
$ docker run -it deepstate bash
user@a17bc44fd259:~/deepstate$ cd build/examples
user@a17bc44fd259:~/deepstate/build/examples$ deepstate-angr ./Runlen
user@a17bc44fd259:~/deepstate/build/examples$ mkdir tmp && deepstate-eclipser ./Runlen -o tmp --timeout 30
user@a17bc44fd259:~/deepstate/build/examples$ cd ../../build_libfuzzer/examples
user@a17bc44fd259:~/deepstate/build_libfuzzer/examples$ ./Runlen_LF -max_total_time=30
user@a17bc44fd259:~/deepstate/build_libfuzzer/examples$ cd ../../build_afl/examples
user@a17bc44fd259:~/deepstate/build_afl/examples$ mkdir foo && echo x > foo/x && mkdir afl_Runlen2
user@a17bc44fd259:~/deepstate/build_afl/examples$ $AFL_HOME/afl-fuzz -i foo -o afl_Runlen -- ./Runlen_AFL --input_test_file @@ --no_fork --abort_on_fail
user@a17bc44fd259:~/deepstate/build_afl/examples$ deepstate-afl -o afl_Runlen2 ./Runlen_AFL --fuzzer_outCheck out docs folder:
DeepState can be used to test R packages written using the popular Rcpp package. The Rcppdeepstate tool is described in a paper presented at the 2021 IEEE International Symposium on Software Reliability Engineering.
All accepted PRs are awarded bounties by Trail of Bits. Join the #deepstate channel on the Empire Hacking Slack to discuss ongoing development and claim bounties. Check the good first issue label for suggested contributions.
We have not yet applied DeepState to many targets, but it was responsible for finding the following confirmed bugs (serious faults are in bold):
- Blosc/c-blosc2#93
- Blosc/c-blosc2#94
- Blosc/c-blosc2#95 (bug causing compression engine to return incorrect uncompressed data) FIXED
DeepState is released under The Apache License 2.0.